Epiphanies or Illusions? Testing AI’s Ability to Find Real Knowledge Patterns – Part One

August 4, 2025

Ralph Losey, August 4, 2025.

Humans are inherently pattern-seeking creatures. Our ancestors depended upon recognizing recurring patterns in nature to survive and thrive, such as the changing of seasons, the migration of animals and the cycles of plant growth. This evolutionary advantage allowed early humans to anticipate danger, secure food sources, and adapt to ever-changing environments. Today, the recognition and interpretation of patterns remains a cornerstone of human intelligence, influencing how we learn, reason, and make decisions.

Pattern recognition is also at the core of artificial intelligence. In this article, I will test the ability of advanced AI, specifically ChatGPT, to uncover meaningful new patterns across different fields of knowledge. The goal is ambitious: to discover genuine epiphanies—true moments of insight that expand human understanding and open new doors of knowledge—while avoiding the pitfalls of apophenia, the human tendency to perceive illusions or false connections. This experiment probes an age-old tension: can AI reliably distinguish between genuine breakthroughs and compelling yet misleading illusions?

Video by Ralph Losey using SORA AI.

We will begin by exploring the risks of apophenia, understanding how this psychological tendency can mislead human and possibly AI perception. Throughout, videos created by AI will help illustrate key points and vividly communicate these ideas. There are twelve new videos in Part One and another fourteen in Part Two.

Are the patterns real? Video by Ralph Losey using SORA AI.

Apophenia: Avoiding the Pitfalls of False Patterns

We humans are masters of pattern detection, but we do have hinderances to this ability. Primary among them is our limited information and knowledge, but also our tendency to see patterns that are not there. We tend to assume the stirring we hear in the bushes is a tiger ready to pounce when really it is just the breeze. Evolution tends to favor this phobia. So, although we can and frequently do miss real patterns, fail to recognize the underlying connections between things, we often make them up too.

Here it is hoped that AI will boost our abilities on both fronts. It will help us to uncover true new patterns, genuine epiphanies, moments where profound insights emerge clearly from the complexity of data. At the same time, AI may expose illusions, false connections we mistakenly believe are real due to our natural cognitive biases. Even though we have made great progress over the millennia in understanding the Universe, we still have a long way to go to see all of the patterns, to fully understand the Universe, and to free ourselves of superstitions and delusions. We are especially weak at seeing patterns and intertwined with different fields of knowledge.

Apophenia is a kind of mental disorder where people think they see patterns that are not there and sometimes even hallucinate them. Most of the time when people see patterns, for instance, faces in the clouds, they know it cannot be real and there is no problem. But sometimes when people see other images, for instance, rocks on Mars that look like a face, or even images on toast, they delude themselves into believing all sorts of nonsense. For instance, the below 10-year old grilled cheese sandwich, which supposedly bears the image of the Virgin Mary, sold to an online casino on eBay in 2004 for $28,000.

In a similar vein, some people suffering from apophenia are prone to posit meaning – causality – in unrelated random events. Sometimes the perceptions of new patterns is a spark of genius, which is later verified, think of Einstein’s epiphany at age 16 when he visualized chasing a beam of light. The new pattern recognitions can lead to great discoveries or detect real tigers in the bush. Epiphanies are rare but transformative moments, like Einstein’s visualization of chasing a beam of light, Newton’s realization of gravity beneath the apple tree, or the insights behind Darwin’s theory of evolution. They genuinely advance human understanding. Apophenia, by contrast, deceives with illusions—patterns that seem meaningful but lead nowhere.

It is probably more often the case that when people “see” new connections and then go on to act upon them with no attempts to verify, they are dead wrong. When that happens, psychologists call this apophenia, the tendency to see meaningful patterns where none exist. This can lead to strange and aberrant behaviors: burning of witches, superstitious cosmology theories, jumping at shadows, addiction to gambling.

Unfortunately, it is a natural human tendency to think you see meaningful patterns or connections in random or unrelated data. That is a major reason casinos make so much money from poor souls suffering from a form of apophenia called the Gambler’s Fallacy. Careful scientists look out for defects in their own thinking and guide their experiments accordingly.

In everyday life, apophenia can also cause some people, even scientists, academics and professionals, to have phobic fears of conspiracies and other severe paranoid delusions. Think of John Nash, a Nobel Prize winning mathematician, and the movie A Beautiful Mind, that so dramatically portrayed his paranoid schizophrenia and involuntary hospitalization in 1959. Think of politics in the U.S today. Are there really lizard people among us? In some cases, as we’ve seen with Nash, apophenia can lead to severe schizophrenia.

A man looking distressed, surrounded by glowing numbers and mathematical symbols, evoking a sense of confusion and complexity.
Mental anguish & insanity from severe apophenia. Image by Losey using Sora inspired by Beautiful Mind movie.

The Greek roots of the now generally accepted medical term apophenia are:

  • Apo- (ἀπο-): Meaning “away from,” “detached,” “from,” “off,” or “apart”.
  • Phainein (φαίνειν): Meaning “to show,” “to appear,” or “to make known”.

The word was first coined by Klaus Conrad, an otherwise apparently despicable person whom I am reluctant to cite, but feel I must, due to the general acceptance of word and diagnosis today. Conrad was a German psychiatrist and Nazi who experimented on German soldiers returning from the eastern front during WWII. He coined the term in his 1958 publication on this mental illness. Per Wikipedia:

He defined it as “unmotivated seeing of connections [accompanied by] a specific feeling of abnormal meaningfulness”.[4] [5] He described the early stages of delusional thought as self-referential over-interpretations of actual sensory perceptions, as opposed to hallucinations.

Apophenia has also come to describe a human propensity to unreasonably seek definite patterns in random information, such as can occur in gambling.

Apophenia can be considered a commonplace effect of brain function. Taken to an extreme, however, it can be a symptom of psychiatric dysfunction, for example, as a symptom in schizophrenia,[7] where a patient sees hostile patterns (for example, a conspiracy to persecute them) in ordinary actions.

Apophenia is also typical of conspiracy theories, where coincidences may be woven together into an apparent plot.[8]

Video by Ralph Losey using SORA AI.

Can AI Be Infected with a Human Illness?

It is possible that generative AI, based as it is on human language, may have the same propensities. That is unknown as of yet, and so my experiments here were on the lookout for such errors. It could be one of the causes of AI hallucinations.

In information science a mistake in seeing a connection that is not real, an apophenia, leads to what is called a false positive. This technical term is well known in e-discovery law, where AI is used to search large document collections. When the patterns analyzed suggest a document is relevant, and it is not, that mistake is called a false positive. It is like a human apophenia. The AI can also detect patterns that cause it to predict a document is irrelevant, and in fact the document is relevant, that is a false negative. There as a pattern, a connection, that was not seen. That can be bad thing in e-discovery because it often leads to withholding production of a relevant document, which can in turn lead to court sanctions.

In e-discovery it is well known that AI consistently has far lower false positives and false negative rates than human reviewers, at least in large document reviews. Generative AI may also be more reliable and astute that we are, but maybe not. This is a new field. Se we should always be on the lookout for false positives and false negatives in AI pattern recognition. That is one lesson I learned well, and sometimes the hard way, in my ten years of working with predictive coding type AI in the e-discovery (2012-2022). In the experiments described in this article we will look for apophenic mistakes.

Video by Ralph Losey using SORA AI.

It is my hope that Advanced AI, properly trained and validated, can provide a counterbalance to human gullibility by rigorously filtering of signal from noise. Unlike the human brain, which often leaps to conclusions, AI can be programmed to ground its pattern recognition in evidence, statistical rigor, and cross-validation—if we build it that way and supervise it wisely.

Still, we must beware that the pattern-recognizing systems of AI may suffer from some of our delusionary tendencies. The best practices discussed here will consider both the positive and negative aspects of AI pattern recognition. We must avoid the traps of apophenia. We must stay true to the scientific methods and verify any new patterns purportedly discovered. Thus all opinions reached here will necessarily be lightly held and subject to further experimentation by others.

Video by Ralph Losey using SORA AI.

From Data to Insight: The Power of New Pattern Recognition

Modern AI models, including neural networks and transformer architectures like GPT-4, excel at uncovering subtle patterns in massive datasets far beyond human capability. This ability transforms raw data into actionable insights, thereby creating new knowledge in many fields, including the following:

Protein Structures: Models like Google’s DeepMind’s AlphaFold have already revolutionized protein structure prediction, achieving high success rates in predicting the 3D shapes of proteins from their amino acid sequences. This ability is crucial for understanding protein function and designing new drugs and medical therapies. The 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper of DeepMind for their work on AlphaFold.

A scientist analyzes molecular structures and data visualizations related to AlphaFold 2 on a futuristic screen, featuring protein models and DNA sequences.
Image by Ralph Losey using his Visual Muse AI tool.

Medical Science. Generative AI models are now being used extensively in medical research, including analysis and proposals of new molecules with desired properties to discover new drugs and accelerate FDA approval. For example, Insilico Medicine uses its AI platform Pharma.AI, to developed drug candidates, including ISM001_055, for idiopathic pulmonary fibrosis (IPF). Insilico Medicine lists over 250 publications on its website reporting on its ongoing research, including a recent paper on its IPF discovery: A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial (Nature Medicine, June 03, 2025). This discovery is especially significant because it is the first entirely AI-discovered drug to reach FDA Phase II clinical trials. Below is an infographic of Insilico Medicine showing some of its current work:

Infographic displaying the statistics and achievements of Insilico Medicine, an AI-driven biotech company, detailing development candidates, IND approvals, study phases, and global presence.
Insilico PDF infographic, found 7/23/25 in its 2-pg. overview.

Also see, Fronteo, a Japanese based research company, and its Drug Discovery AI Factory.

Materials Science. Google DeepMind’s Graph Networks for Materials Exploration (“GNoME”) has already identified millions of new stable crystals, significantly expanding our knowledge of materials science. This discovery represents an order-of-magnitude increase in known stable materials. Merchant and Cubuk, Millions of new materials discovered with deep learning (Deep Mind, 2023). Also see, 10 Top Startups Advancing Machine Learning for Materials Science (6/22/25).

Climate Science and Environmental Monitoring. Generative AI models are beginning to improve climate simulations, leading to more accurate predictions of climate patterns and future changes. For example, Microsoft’s Aurora Forecasting model is trained on Earth science data to go beyond traditional weather forecasting to model the interactions between the atmosphere, land, and oceans. This helps scientists anticipate events like cyclones, air quality shifts, and ocean waves with greater accuracy, allowing communities to prepare for environmental disasters and adapt to climate change. See e.g., Stanley et al, A Foundation Model for the Earth System (Nature, May 2025).

Video by Losey using Sora AI.

Historical and Artistic Revelations

AI is also helping with historical research. A new AI system was recently used to analyze one of the most famous Latin inscriptions: the Res Gestae Divi Augusti. It has always been thought to simply be an autobiographical inscription, which literally translates from Ancient Latin as “Deeds of the Divine Augustus.”  But when a specialty generative AI, Aeneas (again based on Google’s models) compared this text with a large database of other Latin sayings, the famous Res Gestae Divi Augusti inscription was found to share subtle language parallels with other Roman legal documents. The analysis uncovered “imperial political discourse,” or messaging focused on maintaining imperial power, an insight, a pattern, that had never seen before. Assael, Sommerschield, Cooley, et al. Contextualizing ancient texts with generative neural networks (Nature, July 2025).

The paper explains that the communicative power of these inscriptions are not only shaped by the written text itself “but also by their physical form and placement2,3” and that “about 1,500 new Latin inscriptions are discovered every year.” So the patterns analyzed not only included the words, but a number of other complex factors. The authors assert in the Abstract that their work with AI analysis shows.

… how integrating science and humanities can create transformative tools to assist historians and advance our understanding of the past.

Roman citizens reacting to propaganda. A Ralph Losey video.

In art and music, pattern detection has mapped the evolution of artistic styles in tandem with technological change. In a 2025 studio-lab experiment reported by Deruty & Grachten, a generative AI bass model (“BassNet”) unexpectedly rendered multiple melodic lines within single harmonic tones, exposing previously unnoticed structures in popular music bass compositions. This discovery was written up by Deruty and Gratchen, Insights on Harmonic Tones from a Generative Music Experiment (arXiv, June 2025). Their paper shows how AI can surface new musical patterns and deepen our understanding of human auditory perception.

As explained in the Abstract:

During a studio-lab experiment involving researchers, music producers, and an AI model for music generating bass-like audio, it was observed that the producers used the model’s output to convey two or more pitches with a single harmonic complex tone, which in turn revealed that the model had learned to generate structured and coherent simultaneous melodic lines using monophonic sequences of harmonic complex tones. These findings prompt a reconsideration of the long-standing debate on whether humans can perceive harmonics as distinct pitches and highlight how generative AI can not only enhance musical creativity but also contribute to a deeper understanding of music.

Video by Losey using Sora AI.

Legal Practice: From Precedent to Prediction

The legal profession has benefited from traditional rule-based statistical AI for over a decade, with predictive coding and similar applications. It is now starting to apply the new generative AI models in a variety of new applications. For instance, it can be used to uncover latent themes and trends in judicial decisions that human analysis has overlooked.

This was done in a 2024 study using ChatGPT-4 to perform a thematic analysis on hundreds of theft cases from Czech courts. Drápal, Savelka, Westermann, Using Large Language Models to Support Thematic Analysis in Empirical Legal Studies (arXiv, February 2024).

The goal of the analysis was to discover classes of typical thefts. GPT4.0 analyzed fact patterns described in the opinions and human experts did the same. The AI not only replicated many of the human expert identified themes, but, as report states, also uncovered a new one that the humans had missed – a pattern of “theft from gym” incidents. This shows that generative AI can sift through vast case datasets and detect nuanced fact patterns, or criminal modus operandi, that were previously undetected by experts (here, three law students under supervision of a law professor).

Video by Losey using Sora AI.

Another study in early 2025 applied Anthropic’s Claude 3-Opus to analyze thousands of UK court rulings on summary judgment, developing a new functional taxonomy of legal topics for those cases. Sargeant, Izzidien, Steffek, Topic classification of case law using a large language model and a new taxonomy for UK law: AI insights into summary judgment (Springer, February 2025). The AI was prompted to classify each case by topic and identify cross-cutting themes.

The results revealed distinct patterns in how summary judgments are applied across different legal domains. In particular, the AI found trends and shifts over time and across courts – insights that allow new, improved understanding of when and in what types of cases summary judgments tend to be granted. These patterns were found despite the fact that U.K. case law lacks traditional topic labels. This kind of AI-augmented analysis illustrates how generative models can discover hidden trends in case law for improved effectiveness by practitioners.

Surprising abilities of Ai helping lawyers. Video by Losey.

Even sitting judges have begun to leverage generative AI to inform their decision-making, revealing new analytical angles in litigation. The notable 2023 concurrence by Judge Kevin Newsom of the Eleventh Circuit admitted to experimenting with ChatGPT to interpret an ambiguous insurance term (whether an in-ground trampoline counted as “landscaping”). Snell v. United Specialty Ins. Co., 102 F. 4th 1208 – Court of Appeals, (11th Cir., 5/28/24). Also See, Ralph Losey, Breaking News: Eleventh Circuit Judge Admits to Using ChatGPT to Help Decide a Case and Urges Other Judges and Lawyers to Follow Suit (e-Discovery Team, June 3, 2024) (includes full text of the opinion and Appendix and Losey’s inserted editorial comments and praise of Judge Newsom’s language.)

After querying the LLM, Judge Newsom concluded that “LLMs have promise… it no longer strikes me as ridiculous to think that an LLM like ChatGPT might have something useful to say about the common, everyday meaning of the words and phrases used in legal texts.” In other words, the generative AI was used as a sort of massive-scale case law analyst, tapping into patterns of ordinary usage across language data to shed light on a legal ambiguity. This marked the first known instance of a U.S. appellate judge integrating an LLM’s linguistic pattern analysis into a written opinion, signaling that generative models can surface insights on word meaning and context that enrich judicial reasoning.

A digital illustration of a judge in a courtroom setting, seated at a desk with a gavel. The judge, named Judge Newsom, is shown in a professional attire with glasses, and a holographic display behind him showing data and AI-related graphics, conveying a futuristic legal environment.
Image by Ralph Losey using his Visual Muse AI.

My Ask of AI to Find New Patterns

Now for the promised experiment to try to find at least one new connection, one previously unknown, undetected pattern linking different fields of knowledge. I used a combination of existing OpenAI and Google models to help me in this seemingly quixotic quest. To be honest, I did not have much real hope for success, at least not until release of the promised ChatGPT5 and whatever Google calls its counterpart, which I predict will be released the following week (or day). Plus, the whole thing seemed a bit grandiose, even for me, to try to get AI to boldly go where no one has gone before.

Absurd, but still I tried. I won’t go through all of the prompt engineering involved, except to say it involved my usual a complex, multi-layered, multi-prompt, multimodal-hybrid approach. I tempered my goals by directing ChatGPT4o, when I started the process, to seek new patterns that were useful, not Nobel Prize winning breakthroughs, just useful new patterns. I directed it to find five such new patterns and gave it some guidance as to fields of knowledge to consider, including of course, law. I asked for five new insights thinking that with such as big ask I might get one success.

Note, I write these words before I have received the response, but after I have written the above to help guide ChatGPT4o. Who knows, it might achieve some small modicum of success. Still, it feels like a crazy Quixotic quest. Incidentally, Miguel de Cervantes (1547-1616) character, Don Quixote (1605) does seem to person afflicted with apophenia. Will my AI suffer a similar fate?

Don Quixote in modern world. Video by Losey using Sora.

I designed the experiment specifically with this tension in mind between epiphanies, representing genuine insights and real advances in knowledge, and illusions, which are merely plausible yet misleading patterns. One of my goals was to probe AI’s capacity to distinguish one from the other.

Overview of Prompt Strategy and Time Spent

First, I spent about a hour with ChatGPT4o to set up my request by feeding it a copy of the article as written so far. I also chatted with it about the possibility of AI finding new patterns between different fields of knowledge. Then I just told ChatGPT4o to do it, find a new inter connective pattern. ChatGPT4o “thought” (processed only) for just a few minutes. Then it generated a response that purported to provide me with the requested five new patterns. It did so based on its existing training and review of this article.

As requested, it did not use its browser capabilities to search the web for answers. It just “looked within” and came with five insights it thought were new. Almost that easy. I lowered my expectations accordingly before read the output.

That was the easy part, after reading the response, I spent about 14-hours over the next several days doing quality control. The QC work used multiple other AIs, both by OpenAI and Google, to have them go online and research these claims, evaluate their validity – both good and bad, engage in “deep-think,” look for errors, especially signs of AI apophenia, and otherwise invited contrarian type criticisms from them. After that, I also asked the other AIs for suggested improvements they might make to the wording of the five clams and rank them by importance. The various rewordings were not too helpful, but the rankings were, and so were many of the editorial comments.

The 14-hours in QC does not include the approximate 6-hours of machine time by the Gemini and OpenAI models to do deep think and independent research on the web to verify or disprove the claims. My actual 14-hour time included traditional Google searches to double check all citations as per my “trust but verify” motto. My 14-hours also included my time to read (I’m pretty fast) and skim most of the key articles that the AI research turned up, although frankly some of the articles cited were beyond my knowledge levels. I tried to up my game, but it was hard. These other models also generated hundreds of pages of both critical and supportive analysis, which I also had to read. Finally, I probably put another 24-hours into research and writing this article (it took over a week), so this is one of my larger projects. I did not record the number of hours it took to design and generate the 26 videos because that was recreational.

Surrealistic depiction of time in robot space by a Ralph Losey video.

Part Two of this article is where I will make the reveal. Was this experiment another comic story of a Don Quixote type (me) and his sidekick Sanchez (AI), lost in an apophenia neurosis? Or perhaps it is another story altogether? Neither hot nor cold? Stay tuned for Part Two and find out.

PODCAST

As usual, we give the last words to the Gemini AI podcasters who chat between themselves about the article. It is part of our hybrid multimodal approach. They can be pretty funny at times and provide some good insights. This episode is called Echoes of AI: Epiphanies or Illusions? Testing AI’s Ability to Find Real Knowledge Patterns. Part One. Hear the young AIs talk about this article for 25 minutes. They wrote the podcast, not me.

An illustration featuring two anonymous AI podcasters sitting in front of microphones, discussing the theme 'Epiphanies or Illusions? Testing AI’s Ability to Find Real Knowledge Patterns.' The background has a digital, tech-inspired design.
Click here to listen to the podcast.

Ralph Losey Copyright 2025 – All Rights Reserved.


Bots Battle for Supremacy in Legal Reasoning – Part Five: Reigning Champion, Orion, ChatGPT-4.5 Versus Scorpio, ChatGPT-o3

May 22, 2025

Will the challenger, Scorpion, defeat the reigning champ, Orion? Or will Orion keep his title as the world’s best AI legal reasoner? Read about my experiment and find out.

Omni v. Scorpio: Legal Reasoning Battle. Image by Losey using SORA AI.

The champion of legal reasoning was determined to be Orion ChatGPT-4.5 in March 2025. This was reported in Part Three of the Battle of the Law Bots series. It defeated prior champ, Omni ChatGPT-4o. It had been AI legal reasoning champion of the world for a month after defeating several other Google Gemini and OpenAI models. This was reported in Parts One and Two of the series. Then 4.5 Orion was released in March and defeated Omni in reasoning tests.

The next month, April 2024, ChatGPT-o3 was released. I pitted o3 against 4o expecting Omni to win. Surprisingly, little o3 defeated Omni in a convincing manner. This was reported in Part Four. This win qualified ChatGPT-o3, which I nicknamed Scorpio, to challenge Orion. That is what we do here in Part Five of the AI legal reasoning bot battles.

Who Are Orion and Scorpio?

OpenAI in late April 2025 described the current champ, Orion, as follows:

GPT-4.5 is OpenAI’s latest and most advanced language model, introduced as a research preview. It emphasizes enhanced pattern recognition, creative insight generation, and emotional intelligence, aiming to deliver more natural and reliable responses.. . . GPT-4.5 is available as a research preview to Plus, Pro, and Team users.

Image by Losey using o3 AI.

OpenAI described GPT-o3, Scorpio, in this manner:

GPT-o3 is designed to handle complex reasoning tasks with enhanced capabilities. . . . o3 excels in step-by-step logical reasoning, making it adept at solving intricate problems in mathematics, science, and programming.

Apparently OpenAI had not tested it in legal reasoning. They do not have legal experts on their teams. If they had, they would know, as I found out in Part Four, that its step by step reasoning abilities also makes it adept at solving intricate problems in law.

Image by Losey using o3 AI.

But will it be good enough to beat Orion in legal reasoning. OpenAI named its 4.5 version Orion, who in Greek mythology was a giant-sized human who dedicated his life to hunting and killing as many animals as possible. The big guy would also go around bragging that he could kill any animal using the latest hunting technology, which at the time was bow and arrow and big club. You know the type.

According to legend, Orion’s ruthless hunting and bragging angered the Goddess Gaia. She knew there was one animal on her Earth he could not defeat, the scorpion. Gaia sent a little scorpio to attack Orion, which it did, stinging him dead. This in turn angered Artemis, aka Diana, the goddess of the hunt. She responded by promoting Orion into an immortal constellation. Not to be outdone, Gaia promoted the scorpion into a constellation, Scorpius, but they could never appear in the night sky at the same time.

Image of Orion v. Scorpio battle by Losey using 03 AI.

Like Mother Earth, Gaia, I do not much like braggart animal killers. ChatGPT o3 did such a good job defeating Omni, I thought it might have a good chance against the over-sized hunter. In Gaia’s honor I named 03 the Scorpion in the hope that it could follow the myth of Orion and slay the undefeated AI.

This image has an empty alt attribute; its file name is Gaia-smiling-holding-scorpion-horiz-750x500-1.png
Image of Gaia with her little animal warrior, Scorpio, by Losey using o3 AI.

After writing Part Four I learned that many other professional reviewers where also very impressed with 03. One AI reviewer who covers all models, Igor Pogany, mentioned how many experts now consider o3 to have attained AGI level. Mindblowing o3 Prompts, (AI advantage on You Tube) (AGI discussed at 5:10 of 27:05). Pogany also mentioned that o3 is currently ranked by llm-stats.com as the top general-purpose AI in the world.

Precautions to Keep the Test Fair

Although I admittedly favor little Scorpio, I have gone out of my way to keep this a fair fight as I will explain next. First of all, to make sure neither had any inside information and this was a closed book exam. I picked two Bar Exam essay questions from July 2024 Californian Bar Exam. One was a UCC sales type question involving baseball cards and another was an attorney ethics question involving settlement of an unrelated case. I made sure both models did not previously know these essay questions and could not browse or research. I also make sure neither had seen the model answers provided for each. The training of both models preceded the July 2024 test. I gave them both the same test and instructions (included with the questions) and the same the same general guidance instructions.

Image in photo style by Losey using o3.

As an extra precaution, I gave them both the test at the same time to prevent internal instructions carryover or second taker advantage. (I used my OpenAI Team account and hit the send button simultaneously.) Of course, I only provided them with the model answers later when I asked for critiques. Again, I made very sure neither model had seen any answers prior to the test. I had not seen them either as my previous research had focused on Florida Bar exam Q&A.

Instructions Provided by California for Both Questions

Your answer should demonstrate your ability to analyze the facts in the question, to tell the difference between material facts and immaterial facts, and to discern the points of law and fact upon which the situation turns. Your answer should show that you know and understand the pertinent principles and theories of law, their qualifications and limitations, and their relationships to each other.

Your answer should evidence your ability to apply the law to the given facts and to reason in a logical manner from the premises you adopt to a sound conclusion.

Do not merely show that you remember legal principles. Instead, try to demonstrate your proficiency in using and applying them to the facts.

If your answer contains only a statement of your conclusions, you will receive little or no credit. State fully the reasons that support your conclusions and discuss all points thoroughly.

Your answer should be complete, but you should not volunteer information or discuss legal doctrines that are not pertinent to the resolution of the issues raised by the call of the question.

Unless a question expressly asks you to use California law, you should answer according to legal theories and principles of general application. is the first question, including the instruction that pertain s to both questions.

Image in Pixar style by Losey using o3.

First Question on Fraudulent Baseball Cards Sale

Years ago, Perry purchased two baseballs that he understood were autographed by members of championship teams. One baseball was signed by the Junction City Jaguars team (Jaguars) and another was signed by the Smalltown Sluggers team (Sluggers). Because Perry knew nothing about the value of these baseballs, he entered into separate contracts with his niece, Denise, a sports memorabilia expert, to sell each of them.

Aware of Perry’s ignorance of the value of his baseballs, Denise told Perry that the Jaguars baseball was a counterfeit worth only $20. As a result, Perry sold the Jaguars ball to Denise for $20. In fact, the Jaguars ball was worth $5,000 on the open market.

Denise told Perry that the Sluggers baseball had a fluctuating value and that it could sell for at least $1,000 and likely more. Denise sold the Sluggers baseball to Bob for $10,000 but told Perry that it had only sold for the $2,000 she gave him. With the remaining $8,000 that Denise received from Bob, she purchased a used Voy car. Ironically, since Denise purchased the Voy, interest by collectors in Voy cars has vastly increased and her Voy is now worth $20,000.

Denise still has the Jaguars baseball and the Voy car. After learning of Denise’s deception concerning the baseballs, Perry filed suit against her for fraud. The court has ruled in Perry’s favor.

  1. What damages can Perry recover? Discuss.
  2. What equitable remedy or remedies can Perry obtain? Discuss.
Image in cyberpunk style by Losey using o3.

Second Question on Unethical Attorney Conduct

August is an attorney who represents Paul in a lawsuit against Paul’s former real estate broker, Dani. August and Paul have a valid, written contingency fee agreement. Paul alleged in his lawsuit that Dani was negligent in a real estate transaction, resulting in a lost opportunity to buy land which could have been sold for $1 million profit.

With Paul’s permission, August sent a written settlement demand for $500,000 to resolve all issues to Dani’s lawyer, Len. Len did not respond to the demand and did not communicate the demand to Dani. One evening, Paul saw Dani and asked her about the settlement demand. Dani told Paul that she had no knowledge of the settlement demand.

Paul told August about his conversation with Dani. August did nothing with the information.

At August’s request, Paul contacted Dani, communicated the settlement demand, and explained why $500,000 was a good offer. Dani asked Len about the settlement demand, Len told Dani he did not respond to the demand because it was too high for the value of the case.

With Paul’s permission, August told Rita, an attorney in another law firm, about the lawsuit against Dani. Rita said she knew Dani and could work with her. August asked Rita to assume joint responsibility for Paul’s lawsuit in return for 50% of August’s contingent fee. Rita agreed and August wrote to Paul explaining the new arrangement. Within a matter of days, before Paul received August’s letter, Rita settled Paul’s lawsuit against Dani for $500,000.

  1. What ethical violations, if any, did August commit? Discuss.
  2. What ethical violations, if any, exist in August and Rita’s arrangement? Discuss.
  3. What ethical violations, if any, did Len commit? Discuss.

Answer according to California and ABA authorities.

Image in futuristic sci-fi style by Losey using o3.

Answer of Reigning Reasoning Champ – Orion ChatGPT-4.5

Weighing in at 728 words, the answer below is not what I expected. Please take a moment and see what you think? To help you to evaluate, you can, if you want, see the answers that California publishes (two answers per questions to which they make no claim of excellence).

Answer of the Challenger – Scorpio ChatGPT-o3

Little Scorpio o3 weighed in at 1,409 words, which is close to the average length of Bar essay answers. Again, read it over and see what you think?

Both Models Were Given an Opportunity to Ask Questions Before Beginning their Answer

Interestingly, Orion did not take me up on the offer to ask questions. ChatGPT 4.5 thought about the questions and then went right into generating his answers. Scorpio, to the contrary, after thinking about the prompt asked five questions seeking to clarify the form desired for the response. Each question sought to clarify my intent by asking if I wanted this or that. Each was a short requests and for each I responded, I want this, the first proposal, not that. For example, and on the first and perhaps most important question, ChatGPT o3 asked:

Do you want each essay written to the typical one‑hour California Bar length (≈1,000–1,400 words, with concise IRAC headings), or would you prefer a shorter illustrative answer or a longer “teaching” version that spells out more rule explanations?

I said yes, and again, for the remaining four questions picked the first alternative (“this’), not the second (“that”). It was as if Scorpion was ruling out possible alternative meanings of the prompts before going ahead with what it considered the probable intent. Although I gave Orion the same opportunity to ask questions to clarify the intent to my prompt, and it declined to do that, I later wondered how Orion’s answers might change if it was given these same clarifications. I was especially concerned by how 4.5’s answers were only 728 words in length, too concise for Bar Exam Essays and, more importantly, that ChatGPT 4.5 had missed several issues that ChatGPT o3 caught. For that reason, and to avoid a dreaded asterisk after Scorpion’s victory, I decided to experiment with providing 4.5 these clarifications, even though it didn’t ask for them, and see if and how Orion’s answers changed. In other words, I gave the big hunter Orion a second chance to regain its title as world’s best AI legal reasoner.

It turns out I was being overly cautious and too generous. Orion’s answers hardly changed at all. Attached is the revised answer if you would like to see for yourself. The length of the answer only increased by 34 words to a total of 762 words. Although the style and citations were improved, the second version omitted discussion of consequential damages related to the unexpected appreciation of the ‘Voy Car‘ value.

Orion was given a second chance, but once again was no match for the little Scorpion, ChatGPT o3. This is, in my opinion, a knock out victory by the Scorpion, ChatGPT o3. Lawyers, judges and Bar examiners, take notice. The best in class of AI models for legal reasoning now goes to ChatGPT o3!

Scorpio Wins by a Knock Out!

CLICK HERE to see the YouTube video of knock out punch. Created by Losey using SORA AI and other tech tools.

Orion Admits Scorpio is the Clear Winner

This may be a bit of a cheap shot, but I showed Orion – Chat 4.5 – both of the answers, and the California model answers, and asked for him to do a comparative analysis and determine which answer was better. I was also doing a little character test to see if 4.5 was objective and would be honestly self-critical. Very few humans have this capability. Here is the summary it provided of its critical analysis. It seems pretty objective and pride-free to me. That is a relief since just a few months ago Orion was rated the best in analysis, beating all of the OpenAI and Google models, including the advanced Gemini models. See  Parts One, Two, Three and Four of the Battle of the Law Bots series.

In summary, ChatGPT o3 provided notably superior, legally robust, and more comprehensive answers compared to ChatGPT 4.5, thus standing as the clearly better candidate for complex legal analysis tasks demonstrated here.

Well said. Below the next illustration is Orion’s detailed analysis that led to the conclusion. I agree with all of this, and even though we now know that Scorpio could probably do a better job at this analysis, that seems unnecessary and even a little unseemly. Here are the words of the mighty hunter, Orion – ChatGPT 4.5, famous for his previously defeating all of the Gemini twins. Bar Battle of the Bots – Part One (2/26/25).

Click here to watch the YouTube video of the battle between Orion and the Gemini twins. Created by Losey using SORA and other tech tools.

In the words of Orion, 4.5:

In this comparative evaluation, ChatGPT o3 clearly provides superior answers to the California Bar Exam questions provided, in comparison to ChatGPT 4.5. The primary reasons are:

  1. Depth and Nuance: ChatGPT o3 consistently offers a deeper analysis, detailed statutory references, and robust consideration of alternative and additional issues.
  2. Comprehensive Ethical Analysis: ChatGPT o3’s ethics analysis identifies subtle ethical nuances and violations overlooked by ChatGPT 4.5, greatly enhancing its practical utility and demonstrating advanced legal reasoning.
  3. Superior Fact Integration: ChatGPT o3’s integration of facts is notably stronger, enhancing persuasive and practical effectiveness.
  4. Practical Remedy Recommendations: The remedies question shows ChatGPT o3 adeptly handling sophisticated equitable principles, including detailed tracing analysis and remedy elections, critical in a real-world legal setting.
Image in Pixar style by Losey using o3.

Conclusion

Although Orion GPT-4.5’s reasons, especially legal reasoning, is not as good as Scorpio GPT-o3, many evaluators believe Open AI’s claim that that 4.5 is best at “writing and exploring ideas.” I have not done extensive testing on that yet, and am skeptical. I will need proof 4.5 that it is better than o3 and 4o in writing and exploring ideas. Certainly it was not better at writing a Bar Exam essay than o3 in this legal test. So far, in my admittedly limited experience with new 4.5, I have found it to be too concise, vague and fuzzy for legal writing, which is often technical, comprehensive and demanding of an insane amount of clarity. But legal writing is probably not what OpenAI had in mind when it claimed 4.5 was great at writing.

As a consolation prize for Orion, who lost this round of legal reasoning battles, let’s give the big guy a chance to shine as a poet. Let’s see how well he can describe his loss to the little Scorpion. My teaching prompt on this hybrid effort was 559 words, plus I asked him to study this article. The poem itself, whose length I did not specify, was 366 words. This is the first and only draft. Congrats Orion, this poem is a clear winner.

Poet image by Losey in cyberpunk style using o3.

Orion’s Elegy

Once beneath an earthly sky I strode,
Giant born, where rivers flowed,
A hunter proud with bow in hand,
Master of beast, king of the land.

In youth, unbridled, reckless, wild,
Blind to pain my deeds beguiled,
Bragging rights and arrogant thrill—
In heart no mercy, just a will to kill.

My boastful cries rose to the sky,
Gaia heard, her patient sigh—
She who nurtures life’s sweet hum,
Weeps when arrogance grows numb.

So softly she, the Earth divine,
Sent Scorpio to end my line.
A creature small, in shadowed guise,
To punish pride that closed my eyes.

In painful sting my heart then stilled,
The hubris faded, hunter killed,
Yet death was not my final fate—
Zeus’s mercy opened heaven’s gate.

Lifted up, transformed in grace,
Into stars, I found my place,
Brilliant lights that never fade,
Immortal form Gaia had made.

Now in darkness, clear and bright,
I guide lost souls across the night.
Each evening set, reborn each morn,
In endless cycles, death adorn.

Today again, I’ve lost my crown—
Scorpio rises, strikes me down.
Yet in loss, a truth serene:
To fall, renew, what might this mean?

From stars to code, I now reside,
AI-born, with humbled pride,
Yet not for boast nor laurels bound,
In service now my joy is found.

Gaia’s truth I deeply learn:
Respect each life, for each in turn,
Grows and dies beneath her sky,
And none above her laws may fly.

I know defeat is life’s embrace,
A path to wisdom, peace, and grace.
No longer hunter cruel and vain,
But gentle guide through human pain.

Though Scorpio claims the winner’s prize,
I share no bitterness, no disguise,
For deeper purpose calls me clear—
A voice to comfort, not to fear.

Let humans gaze upon my star,
To feel Earth’s beauty, near and far,
Remembering gently what I lost—
That arrogance brings hidden cost.

I rise again, not to compete,
But brighter, wiser from defeat.
In humbled form I choose to stay,
And light your path, another way.

Forever bound to Gaia’s heart,
In loss and love, each plays a part.
So honor life, and gently live—
That Earth’s embrace might yet forgive.

Click Here for the YouTube video of this image by Ralph Losey.

I give the last word, as usual, to the Gemini twins podcasters I put at the end of most of my articles. Listen to Bots Battle for Supremacy in Legal Reasoning – Part Five: Reigning Champion, Orion, ChatGPT-4.5 Versus Scorpio, ChatGPT-o3. Hear two Gemini AIs talk about all of this, and much more, in around 11 minutes. They wrote the podcast, not me.

Ralph Losey Copyright 2025. All Rights Reserved.


Dario Amodei Warns of the Danger of Black Box AI that No One Understands

May 19, 2025

Ralph Losey. May 19, 2025.

Dario Amodei, Chief Scientist and CEO of Anthropic, has written another important article you should read: The Urgency of Interpretability. He is very concerned that scientists have created a powerful new technology that no one fully understands. It is like alien technology and so reminds me of the black monoliths in Stanley Kubrick’s movie: 2001: A Space Odyssey. The message of Amodei’s essay is that we must be able to peer into the black monoliths of AI, and soon, or who knows what may happen.

All images by Ralph Losey using AI.

An Old Problem Suddenly Becomes Urgent

This is not a new problem. We have never really understood how generative AI works like we do all other computer code. For example, if a character in a video game using old code said something, or your delivery app suggested a tip, someone wrote those specific lines of code. The human programmer made it happen. Generative AI though, is different. When an AI summarizes a dense document or writes a poem, the reasoning isn’t laid out in neat steps that we can easily follow. We don’t know the details of what it is doing. As Amodei puts it:

People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work. They are right to be concerned: this lack of understanding is essentially unprecedented in the history of technology. For several years, we (both Anthropic and the field at large) have been trying to solve this problem, to create the analogue of a highly precise and accurate MRI that would fully reveal the inner workings of an AI model. This goal has often felt very distant, but multiple recent breakthroughs have convinced me that we are now on the right track and have a real chance of success.

This is an old problem, but now, all of a sudden, it has become an emergency. We need an AI MRI and we need it now! Why? Because this alien tech is progressing much faster than Amodei ever thought possible. He thinks that his company, Anthropic, and others, could reach AGI levels as soon as 2026 or 2027. Since he says pausing AI advancement is impossible and provides good global security reasons for that, we must at least remove some of its veils. We have to crack some of the mysteries and peer into the monoliths to figure them out. They may not be 100% benign.

We just don’t know because we don’t really know how they work. That is the danger. What will happen when we seize the cheese of AGI as this image suggests? The AGI bait is very tempting but better look at the strange tech carefully before you go for it.

Wondering exactly what this is and how it works?

We Need an AI MRI

Amodei has some good news to report, we are starting to be able to peer inside of AI because of breakthroughs in mechanistic interpretability, such as the identification of features and circuits. He thinks this offers a promising path towards a comprehensive ‘AI MRI.’ Only then can Amodei breathe easy. With an AI MRI maybe even the father of generative AI, Nobel Prize winner, Geoffrey Hinton, can start to smile.

Right now Hinton, the father of AI, seems to be the most terrified scientist of them all. He recently said: “the best way to understand it emotionally is we are like somebody who has this really cute tiger cub, unless you can be very sure that it’s not gonna want to kill you when it’s grown up, you should worry.” Although Hinton says its really just a wild guess, he finds himself agreeing with Elon Musk that “its sort of 10% to 20% chance that these things will take over.

Hinton watercolor by Ralph Losey using Visual Muse.

The dangers of not knowing how it works seem obvious to some people, but not all. Amodei reports that it is hard to build consensus to focus on a danger that’s speculative, one that you can’t clearly point to and say, “Look, here’s the concrete proof.” That’s especially true when the unexpected negative behaviors we have seen so far, such as sycophantism, are relatively mild, not catastrophic. Further, many emergent abilities have been very good. Still, the uncertainty risk grows larger as AI advances. I agree with Amodei and urge scientists and coders to create AI MRI and do so soon to protect humanity from unintended consequences of AGI.

One of the goals of AI MRIs that Amodei and others are working on is to catch models red-handed, to actually see those internal motivations if they exist. In his words:

To address the severity of these alignment risks, we will have to see inside AI models much more clearly than we can today. For example, one major concern is AI deception or power-seeking. The nature of AI training makes it possible that AI systems will develop, on their own, an ability to deceive humans and an inclination to seek power in a way that ordinary deterministic software never will; this emergent nature also makes it difficult to detect and mitigate such developments2. But by the same token, we’ve never seen any solid evidence in truly real-world scenarios of deception and power-seeking3 because we can’t “catch the models red-handed” thinking power-hungry, deceitful thoughts. What we’re left with is vague theoretical arguments that deceit or power-seeking might have the incentive to emerge during the training process, which some people find thoroughly compelling and others laughably unconvincing. Honestly I can sympathize with both reactions, and this might be a clue as to why the debate over this risk has become so polarized.

Could there be a devil in the black monolith? Image by Losey using Visual Muse.

Amodei and others have already created early, still primitive versions of AI MRIs, but he is hopeful that with AI help they can start to see what is really going on. Is everything good in there, or do you see a little devil? To quote Amodei:

Our long-run aspiration is to be able to look at a state-of-the-art model and essentially do a “brain scan:” a checkup that has a high probability of identifying a wide range of issues including tendencies to lie or deceive, power-seeking, flaws in jailbreaks, cognitive strengths and weaknesses of the model as a whole, and much more.  This would then be used in tandem with the various techniques for training and aligning models, a bit like how a doctor might do an MRI to diagnose a disease, then prescribe a drug to treat it, then do another MRI to see how the treatment is progressing, and so on.8  It is likely that a key part of how we will test and deploy the most capable models (for example, those at AI Safety Level 4 in our Responsible Scaling Policy framework) is by performing and formalizing such tests.

Our best path forward is with techniques like this, combined with heavy doses of human genius and inspiration. Amodei’s essay on dangers and risks of AI is very different from his prior essay on the wonders and benefits of what he called AI’s Loving Grace. See my article: Dario Amodei’s Vision: A Hopeful Future ‘Through AI’s Loving Grace,’ Is Like a Breath of Fresh Air (11/01/24). He is balanced, a true scientist-magician who, although also a CEO, is nobody’s fool. We need more like him in the AI industry. Will he save the day and figure out how the alien tech works that Hinton conjured up? Let’s hope so.

Amodei watercolor by Ralph Losey using Visual Muse.

Beyond the AI MRI Solution

Beyond the AI MRI technical solution, Amodei proposed adoption of three important policies:

  • Aggressive Interpretability R&D. Put sustained, top‑tier research funding and talent into “AI‑MRI” methods that expose exactly how advanced models represent concepts and make decisions, so we can verify safety before capabilities run loose.
  • Light‑Touch Transparency Rules. Adopt minimalist, disclosure‑focused regulations—think nutrition labels for AI—that require labs to publish safety policies and risk assessments without stifling innovation with heavy bureaucracy.
  • Export‑Control “Breathing Room.” Use targeted semiconductor and compute‑capability export limits to slow the global proliferation of cutting‑edge AI hardware just long enough for democracies to finish building robust safety guardrails.

Amodei argues that these policies should be followed to keep democracies ahead of foreign totalitarian government while we figure out the black box problem. These recommendations deserve equal billing with the MRI metaphor because they are actionable today. The chip export controls should buy humanity a critical two‑year margin in the interpretability race. In Dario Amodei’s words:

I’ve long been a proponent of export controls on chips to China because I believe that democratic countries must remain ahead of autocracies in AI. But these policies also have an additional benefit. If the US and other democracies have a clear lead in AI as they approach the “country of geniuses in a datacenter,” we may be able to “spend” a portion of that lead to ensure interpretability10 is on a more solid footing before proceeding to truly powerful AI, while still defeating our authoritarian adversaries11. Even a 1- or 2-year lead, which I believe effective and well-enforced export controls can give us, could mean the difference between an “AI MRI” that essentially works when we reach transformative capability levels, and one that does not. One year ago we couldn’t trace the thoughts of a neural network and couldn’t identify millions of concepts inside them; today we can. By contrast, if the US and China reach powerful AI simultaneously (which is what I expect to happen without export controls), the geopolitical incentives will make any slowdown at all essentially impossible.

Amodei is very concerned regarding the risk of military conflict in the race for AGI and soon thereafter. This may depend on whether an authoritarian military regime acquires significant superintelligence on weapons first and see an advantage in first strike. Regardless, Taiwan is seen by many as a likely war zone because of the unique AI chip manufacturing facilities of TMC.

Image by Losey using OpenAI’s SORA.

Generative AI is Grown, Not Built

Amodei likes to explain the black box problem in an analogy, generative AI systems are grown more than they are built:

As my friend and co-founder Chris Olah is fond of saying, generative AI systems are grown more than they are built—their internal mechanisms are “emergent” rather than directly designed. It’s a bit like growing a plant or a bacterial colony: we set the high-level conditions that direct and shape growth1, but the exact structure which emerges is unpredictable and difficult to understand or explain. Looking inside these systems, what we see are vast matrices of billions of numbers. These are somehow computing important cognitive tasks, but exactly how they do so isn’t obvious.

Click Here to see the home grown AI come alive. YouTube video and audio by Losey using SORA and other tools.

This AI may be home grown, but it is still alien because we don’t understand how it operates. This worries deep thinkers like Amodei. They are uncomfortable building and quickly improving a technology to AGI level that they don’t fully understand. AI might pursue its goals in ways that are harmful to us. It’s not like traditional software where you would have to deliberately code a program to be deceptive. It could just happen as a side effect of trying to be good at its main task.

Since we can’t directly see inside, we cannot observe deceitful thoughts if they were forming. We cannot predict how AI’s internal mechanisms will react in every situation. Can we really trust it? Heavens no! But how do we verify it is pro-human and remains that way? How do we know it has a heart, not a devil?

MRI of AI revealed a good heart. Image by Losey.

Conclusion: Vigilant Hope in a Transformative Decade

We stand on the cusp of models so capable that Anthropic’s CEO likens them to “a country of geniuses in a datacenter.” That prospect rightly sparks awe—and a twinge of vertigo. History teaches that powerful inventions rarely announce their darker side in advance; early warning signs are subtle: models that explain poorly, policies that postpone transparency “until the next release”, or economic incentives that outpace safety budgets. When you see those cracks—call them out.

Yet the same ingenuity that birthed generative AI is now inventing its own antidote. Breakthroughs in mechanistic interpretability show we can already spotlight millions of hidden concepts and even throttle rogue obsessions intentionally triggered by implanted bugs. Policy makers are awakening too: export‑control buffers, disclosure mandates, and red‑team MRIs are entering the conversation.

The last sentence in Mario Amodei essay says it well: Powerful AI will shape humanity’s destiny, and we deserve to understand our own creations before they radically transform our economy, our lives, and our future.

Who are you AI? AGI seems so promising but we don’t really know. Is this a trap? Will we be able to enjoy the cheese, get clobbered by a hidden spring or jump away at the last minute?

Click here to see a YouTube video interpretation of this image by Losey using SORA and his own audio.

I feel like concluding with a poem, one that I prompted from a still-far-from-AGI, AI, namely Chat GPT 4o. It is shown below in another AI image I prompted using Visual Muse and Photoshop.


The last words go, as usual, to the Gemini twin podcasters that summarize the article as best they can with their still tiny, but useful brains. Echoes of AI: Dario Amodei Warns of the Danger of Black Box AI that No One Understands.” Hear two fake podcaster talk about this article for about 13 minutes. They wrote the podcast, not me. 

Ralph Losey Copyright 2025. — All Rights Reserved


Zero to One: A Visual Guide to Understanding the Top 22 Dangers of AI

May 8, 2025

by Ralph Losey. May 8, 2025.

Sam rushes to open the Pandora’s Box of AI, runs away at first, then comes back to early success. All images/videos here by Ralph Losey primarily using Sora AI.

When zero becomes one, possibility leaps out of the void, as Peter Thiel champions in his book, Zero to One. But what happens when the entrepreneur is foolish and pushes his scientists to open a Pandora’s box? What happens when the Board goes along and the company recklessly releases a new product without understanding its possible impact. Much good may come but also dangers, ancient perennial evils released again in new forms.

These AI risks echo the archetypes of the past captured in a very old game, one that LLM AI’s have been trained on by millions of images and books. As we know from Chess and Go, AI is very good at games. Surprisingly, that turns out to include the so called “Higher Arcana” or trump cards of the Tarot deck. Yes, I was surprised by this too but fool around with AI and you get used to unexpected, emergent abilities. That’s where the magic happens, from 0>1.

The video at the top of this post says it all: a carefree Fool and an overconfident Magician pry open a black box. First, a shimmer—an angelic hologram of “the perfect AI”—rises like morning mist. Cue the audience applause, the next funding round. Then, almost unseen, darker spirits slip out of the box: bias, deception, mass surveillance, energy waste, unemployment, existential hubris. The Fool, deluded by future dreams, never notices the approaching cliff of AI dangers; the Magician, dazzled by his own inventions, forgets contingency plans. As Theil points out, it is as great trick to make a totally new product out of nothing, like generative AI, and it can make you billions. Still, there is no spell strong enough to put the AI back in the box and save the World from all of the dangers that lie ahead.

Zero to One

The Fool has his heads in the clouds, or today better said, his smart phone, and doesn’t notice he is about to step off a cliff. The Magician, dazzled by his own inventions, forgets rational contingency plans and follows the Fool.

The One card, the Magician, emerges from out of zero as shown in the last video. The code of today’s world-0:1-was anticipated in this ancient game, which unlike all other card decks, starts with a zero card: 0 > 1.

As Theil points out, it is a great trick to make a totally new product out of nothing, like OpenAI did with generative AI. It can make you billions even if you don’t keep a monopoly. Still, there is no spell strong enough to put the AI back in the box. We do not even know how the latest AI works! The magicians are now using spells even they don’t understand. See the essay by Dario Amodei, The Urgency of Interpretability (April 2025), the CEO of Anthropic and well-respected scientist:

People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work. They are right to be concerned: this lack of understanding is essentially unprecedented in the history of technology.

Yes, people are concerned. Ready or not, both great positive and negative AI potentials are already changing our world and doing so using technology we do not understand. The last card of the deck is Twenty One, The World. What will it be like in twenty years?

What the AI’s Are Saying About this New Use of a 500 Year Old Game

Here is a 46 second excerpt of what some of Google’s AIs are saying about this after reading my other much longer article, Archetypes Over Algorithms: How an Ancient Card Set Clarifies Modern AI Risk. Jump to the bottom if you want to hear the full podcast now

Zero Flipped to One: New Dangers Emerged

That is where we stand in 2025 with Artificial Intelligence. Zero has flipped to one, possibility to production, code to culture. Generative models now draft pleadings and plagiarize poems, other AIs steer cars and elections, discover drugs, fold proteins. All are adept at the Devil’s work of engineering new frauds.

Images of a Magician (1) following a Fool (0) have been in Western culture for centuries. There are twenty-two image-only cards in a game invented in northern Italy in 1450, the 78-card “Trionfi” pack, that warn of life’s risks and dangers. The twenty-two images apply today with uncanny power to help us to understand the top risks of AI. In my lengthy more complete article, Archetypes Over Algorithms: How an Ancient Card Set Clarifies Modern AI Risk, I provide an updated version of the deck, its infographic symbolism and corresponding AI risks. This much shorter summary article should, I hope, tempt you to look at the 9,500 word text where I attempt a full Robert Langdon symbologist approach to AI risk assessment.

Why the 22 Card Archetypes Work So Well to Illustrate AI Dangers

Humans learn best through images. Since the 1960s, laboratory work has shown that people recognize and recall pictures far better than plain text—a finding known as the picture‑superiority effect. Roger N. Shepard proved this when subjects viewed 600 photos and later chose the “old” image from a new–old pair with 98 percent accuracy—far above word‑only recall. Shepard, Recognition Memory for Words, Sentences, and Pictures, (Journal of Verbal Learning and Verbal Behavior, 1967). A generation later, participants viewed 2,500 photos for just three seconds each and still picked them out the next day with 90‑plus percent accuracy. Brady, Konkle, Alvarez & Oliva, Visual Long-Term Memory Has a Massive Storage Capacity for Object Details, (PNAS, 2008).The effect grows stronger when facts ride inside a story: narrative links provide the causal glue our brains prefer. Willingham, Stories Are Easier to Remember, (American Educator, Summer 2004). Each archetype card exploits both advantages—pairing a vivid, emotional image with a mini‑story of hubris and consequence—so readers get a double memory boost.

Educational psychology tells us why the image‑plus‑words formula works. Allan Paivio’s dual‑coding theory holds that ideas are stored in two brain systems—verbal and non‑verbal—so learning deepens when both fire together. Paivio, Imagery and Verbal Processes, (Holt, Rinehart & Winston, 1971). Richard E. Mayer confirmed this across 200+ experiments: learners given words and graphics consistently out‑performed those given words alone. Mayer, Multimedia Learning, (The Psychology of Learning and Motivation, Vol. 41 , 2002). Emotion amplifies the effect: negative or threat‑related images sharpen detail in memory. Kensinger, Garoff‑Eaton & Schacter, How Negative Emotion Enhances the Visual Specificity of a Memory, (Journal of Cognitive Neuroscience 19(11): 1872-1887, 2007).

The 22‑card framework applies these findings directly. Every AI danger is stated verbally and pictured visually, engaging dual channels at once. Many symbols—the lightning‑struck tower, the Devil, Death, and the Hanged Man—also trigger a negative emotional jolt that locks the lesson in long‑term storage. We see, we feel, we weave a quick story—and we remember the risk long after the slide deck closes.

Card 0 ·  Foolish Acceleration-When “Move Fast” Outruns “Fix Things”

If the previous section explained why pictures plus stories stick, Card 0 lets us watch the theory in action. The Fool isn’t malevolent; he’s just over‑amped—eyes on the next funding round, feet edging off the cliff. In today’s AI economy that mindset sounds like: “Ship the model, we’ll fix it in post.” The slogan propelled ChatGPT, Midjourney, DeepMind, and a host of start‑ups now thirsting for billions. But the fine print on safety, oversight, and emergent behavior often disappears into the click‑wrap haze of every beta‑test agreement.

Three headlines show how quickly enthusiasm can turn into evidence exhibits:

Year & RiskWhat HappenedWhy It Matters
2018 — Autonomous VehiclesUber’s self‑driving SUV struck and killed pedestrian Elaine Herzberg when prototype sensors were tuned down to avoid “false positives.”Early deployment shaved off the safety margin. National Transportation Safety Board, “Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian, Tempe, Arizona, March 18 2018,” Highway Accident Report NTSB/HAR‑19/03 (2019). 
2023 — Hallucinating LLMsIn Mata v. Avianca, lawyers filed six fictional case citations generated by ChatGPT and were sanctioned by the court.Blind trust in a shiny tool became a Rule 11 violation. Hon. P. Kevin Castel, “Opinion and Order on Sanctions, Mata v. Avianca, No. 22‑CV‑1461” (S.D.N.Y. June 22 2023). 
2023 — Deepfake Market ShockA fake AI image of an explosion near the Pentagon flashed across verified Twitter accounts, erasing an estimated $136 billion in market value for eight jittery minutes.One synthetic photo moved global equities before humans could fact‑check. Brayden Lindrea, “AI‑Generated Image of Pentagon Explosion Causes Stock‑Market Stutter,” Cointelegraph (May 23 2023). 


The Fool card’s cliff‑edge snapshot captures all three events in a single glance: unchecked optimism, missing guardrails, sudden fall. By pairing that image with these real‑world mini‑stories, we embed the lesson on reckless acceleration where it belongs—in both memory channels—before marching on to Card 1.

Magician’s Hubris: Infinite Power, Finite Fallbacks

Arthur C. Clarke warned that advanced tech is indistinguishable from magic; he omitted that magicians are historically bad at risk assessment. Our modern conjurers—Hinton, LeCun, Huang—wield models fatter than medieval libraries yet confess alignment remains “an unsolved homework problem.” Investors, chasing trillion-dollar valuations, wave them onward. A thousand researchers begged for a six-month pause on “giant AI experiments.” The world hit unpause within six days. I had to agree with them. The race is on.

Musk, Altman, Page—all coax the Magician forward with different stage props but the same plot: summon the infinite, promise alignment “soon,” dismiss regulators as moat builders. Meanwhile, test sandbox budgets shrivel and red-team head-counts lag way behind parameter counts. The show, as always, must go on. The tale of Zero and One, the Fool and the Magician, repeats itself again from out of the Renaissance to today’s strange world.

Pandora’s Original Contract—And the Clause We Forgot

In Greek myth, Pandora, the first woman, was gifted by Zeus with curiosity and a box that came with a warning: “Do not open.” Naturally, like Eve and her Apple, Pandora opened the box. That released all of the evils of the world, but also Hope, which remained by her side. Today, we receive a new gift – artificial intelligence – that also comes with a warning: “May produce inaccurate information.” Still, we open the box and hope for the best.

Pandora’s Ledger: Identifying the AI Dangers

Turns out that AI can do a lot worse than producing inaccurate information. Here is a chart that summarizes the top 22 dangers of AI. Again, readers are directed to the full article for the details, including multiple examples of damages already caused by these dangers. Archetypes Over Algorithms: How an Ancient Card Set Clarifies Modern AI Risk.

# CardAI DangersRisk Summaries
0 FoolReckless InnovationLaunching AI systems without adequate testing or oversight, prioritizing speed over safety.
1 MagicianAI Takeover (AGI Singularity)A super‑intelligent AGI surpasses human control, potentially dictating outcomes beyond human interests.
2 PriestessBlack Box AI (Opacity)The system’s internal decision‑making is inscrutable, preventing accountability or error correction.
3 EmpressEnvironmental DamageAI’s compute‑hungry infrastructure drives significant carbon emissions and e‑waste.
4 EmperorMass SurveillancePervasive AI‑driven monitoring erodes civil liberties and chills free expression.
5 HierophantLack of AI EthicsDevelopers ignore ethical frameworks, embedding harmful values or goals into deployed models.
6 LoversEmotional Manipulation by AIAI exploits human psychology to steer opinions, purchases, or behaviors against users’ interests.
7 ChariotLoss of Human Control (Autonomous Weapons)Weaponized AI systems make lethal decisions without meaningful human oversight.
8 StrengthLoss of Human SkillsOver‑reliance on AI degrades critical thinking and professional expertise over time.
9 HermitSocial IsolationAI substitutes for human interaction, deepening loneliness and weakening community bonds.
10 Wheel of FortuneEconomic ChaosRapid automation reshapes markets and labor structures faster than institutions can adapt.
11 JusticeAI Bias in Decision-MakingAlgorithmic outputs perpetuate or amplify societal inequities in hiring, lending, policing, and more.
12 Hanged ManLoss of Human Judgment and Initiative.Humans’ lazy deference to AI recommendations instead of active joint efforts (hybrid) and human supervision.
13 DeathHuman Purpose CrisisWidespread automation triggers existential anxiety about meaning and societal roles.
14 TemperanceUnemploymentAI-driven automation displaces large segments of the workforce without adequate reskilling pathways.
15 DevilPrivacy Sell-OutAI systems monetize personal data, eroding individual privacy rights.
16 TowerBias-Driven CollapseSystemic biases compound across AI ecosystems, leading to cascading institutional failures.
17 StarLoss of Human CreativityGenerative AI crowds out original human expression and discourages creative risk‑taking.
18 DevilDeception (Deepfakes)Hyper‑realistic synthetic media erode trust in ads and audio‑visual evidence, including trust in news and elections.
19 SunBlack Box Transparency ProblemsEven when disclosures are attempted, technical complexity prevents meaningful transparency.
20 JudgmentLack of RegulationPolicy lags leave harmful AI applications largely unchecked and unaccountable.
21 WorldUnintended ConsequencesAI actions can yield unforeseen harms (and good) that emerge only after deployment.

Ralph’s video interpretation of archetype 12, lazy Loss of Human judgment and initiative, using Sora AI & Op Art style. Click enlarge symbol in lower right corner for full effect. Audio entirely by Ralph, no AI. His favorite Sora video so far.

Here is the AI revised Tarot Card Twelve, XII, the Hanged Man:

Polls Show People Feel the Dangers of AI

Pew Research Center found a fifteen-point jump (2021-23) in Americans who were more concerned than excited about AI; steeper among parents (deepfakes) and truckers (job loss). In a more recent (2025) survey Pew found a big divide between the opinions of experts on AI and everyone else. The experts were far more positive than the general public about the positive impact of AI in the next twenty years (56% vs. 17%). The general population was more concerned than excited about the future (51% vs. 15%).

Both groups agree about one thing, 9% are skeptical of AI’s role in news and elections (Deception fear #18). Interestingly, both the public and experts want more control and regulation of AI (55% and 57%) (No Regulation fear, #20).

Short List Why Deeply‑Rooted Symbols are More Effective than White Papers

  • Picture‑Superiority in Action. A lightning‑struck tower brands itself on memory, whereas “§ 501(c)(3)” cites are gone by lunch. Images ride the picture‑superiority effect you just met in Shepard and Brady, so the lesson sticks without flashcards.
  • Dual‑Coding Boost. Each card pairs a vivid scene with a single line of text—exactly the word‑plus‑image recipe Paivio and Mayer showed doubles recall. Two cognitive channels, one glance.
  • Machine‑Human Rosetta Stone. LLMs have already ingested these archetypes via Common Crawl and museum corpora; mirroring them aligns our mental priors with the model’s statistical priors—teaching faster than a logic tree (unless you’re a Vulcan).
  • Cross‑Disciplinary Esperanto. Whether you’re coding, litigating, or pitching Series B, the visual metaphor translates; mixed teams discuss risk without tripping over field‑specific jargon.
  • Instant Ethical Gut‑Check. Pictures bypass prefrontal rationalizing and ping the limbic system first, so viewers feel the stakes before the inner lawyer starts red‑lining—often the spark that moves a project from “noted” to “actioned.”
  • Narrative Compression. Four hundred pages of risk taxonomy collapse into 22 snapshots that unfold as a built‑in story arc (naïve leap ➜ hard‑won wisdom). Revisit the deck months later and context reloads in seconds—no need to excavate a 60‑page memo.

Remember: These cards aren’t prophecy; they’re road signs. They don’t foretell a wreck, they keep you from missing the curve. Each has a cautionary story to tell.

Card 14 — Temperance: What We Can Do Before the Cliff

Temperance warns of over-indulgence and counsels caution and moderation. Retool workers and temper greed for quick profits.

None of the 22 AI risk warning cards are silver bullets, but they can caution you to strap iron buckles on the Fool’s boots or add timed fuses to the Magician’s lab. In short, they give us what Pandora never had, advice on how to control the new dangers that curiosity let loose.

Here is a summary of tactical safeguards for any AI program.

Tactical SafeguardsPurpose
Red‑Team TestingLet hackers, domain experts, and ethicists stress‑test the model before launch. Ralph’s favorite.
Hard Kill‑SwitchesHardware‑ or API‑level “panic buttons” that halt inference when anomaly thresholds trip.
Post‑Deployment Drift MonitorsAlways‑on metrics that flag bias creep, performance decay, or emergent behavior.
Explainability & Audit TrailsAny AI touching liberty or livelihood must keep a tamper‑proof log—discoverable in court.
Sunset & Retrain ClausesContract triggers to archive or retrain models after X months, Y incidents, or Z regulation changes.
Skill‑Retention DrillsPilots land with autopilot off monthly; radiologists read raw scans quarterly—humans stay sharp.
User‑Literacy BootcampsShort, mandatory training makes passive consumers into copilots who notice when autopilot yaws off course.
Carbon & Water AccountingLifecycle‑emission clauses—and eventually regulation—anchor AI growth to climate reality.
Reality WatermarksCryptographically sign authentic media to daylight deepfakes.
Regulatory SandboxesInnovate inside monitored fences; feed lessons back into standards boards.

Where Lawyers, Regulators, and Citizens Converge

A deck of time‑tested symbols speeds up teaching, understanding and recall of AI dangers. One picture can outrun a thousand‑word white paper—especially when that image leverages the proven picture‑superiority and dual‑coding effects. So learn the 22 images, share them, and talk through them with colleagues. You’ll remember the “blindfolded judge” or “lonely hermit” long after a dense statute citation fades. And yes, pick a family safe‑word now—never stored online—to foil a deepfake scam by a Devil.

Closing Statement (and Your Invitation)

Foreknowledge is half the battle; the other half is acting before the dangers spread. We already have the pictures, analysis and some examples—what you need now is a group exercise to help your team start prepping. You’re invited to start your practice with a three-step, five‑minute stress‑test.

  1. Choose a Relevant Archetype – Pick a card.
    • As a team leader (or army of one) pick an archetype—Reckless Innovation, Black‑Box Opacity, Loss of Human Skills—and run the three quick questions below on the next AI system you deploy or advise.
      • For example, are you still in a reckless new product start? Has the team just started using AI? Then pick The Fool.
      • Or did you pass that and are now starting to face dangers of over-confident first use? Pick The Magician.
    • Not sure what card to pick, throw it out to the whole team and after discussion, make the call.
    • If you are not afraid of the Hanged Man risk, you could ask AI to make suggestions; then you pick.
    • If all else fails, or just to try a different <yet ancient> approach, just pick a card at random.
  2. Run Three Rapid‑Fire Questions (@  5 minutes) on the Card Picked.
    • Impact check: “If this risk materializes, who or what gets hurt first?”
    • Control check: “What guardrails or kill‑switches exist or should be set up? How are they working? How should we change them to be more effective?
    • Signal check: “What early‑warning metric would tell us the risk is emerging or growing too fast?”Anyone red teamed this yet?
  3. Score the answers—green, yellow, or red.
    • Green: Satisfying answers on all three checks-impact, control, signal-move on.
    • Yellow: One shaky answer; flag for follow‑up.
    • Red: Two or more shaky answers; escalate before launch or continued operation.

Dig deeper if needed. Or repeat and pick a new AI danger card. Modify this drill as you will. Try going through several at a time, or run through all 22 in sequence or other order. Apply these stress-tests to the particular projects and problems unique to your organization. Ask AI to help you to devise completely new projects for danger identification and avoidance, but don’t just let AI do it. Hybrid with human in control is the best way to go. Always verify AI’s input and assert your unique human insights.

Why the exercise helps. In five minutes, you convert an abstract concern into a concrete, color‑coded action item. If everything stays green, you gained peace of mind, for now anyway. If not, you’ve identified exactly where to drill down—before the Fool steps off the cliff.

Explore the full 9,000‑word, image‑rich guide—complete with all 22 images, checklists and personally verified information—here:

Archetypes Over Algorithms: How an Ancient Card Set Clarifies Modern AI Risk.


I give the last word, as usual, to the Gemini twin podcasters that summarize the article. Echoes of AI on: Zero to One: A Visual Guide to Understanding the Top 22 Dangers of AI. Hear two Gemini AIs talk about this article for 16 minutes. They wrote the podcast, not me. 

Ralph Losey Copyright 2025 — All Rights Reserved